MACHINE LEARNING PROCEDURES FOR GENERATING IMAGE DOMAIN FEATURE DETECTORS

This dissertation presents a machine learning system for generating image domain feature detectors. The feature detectors are programs for a cellular image processor employing the operators of mathematical morphology (dilation, erosion, union, complement, and intersection). The learning system uses genetic search to generate populations of feature detectors which cooperate in the classification of image samples. In a first series of experiments, the system is shown digitized images of text samples from different type scripts and generates feature detectors which allow the system to identify which type script a given text sample comes from. The same system is then used to generate feature detectors for classifying cartoon faces (eg. Charlie Brown vs. Snoopy). In these experiments the system classified both training and test samples with 100% accuracy. The dissertation addresses several theoretical issues concerning knowledge representation, computer vision and machine learning. Within the field of computer vision the fundamental question of what primitives should be extracted from the image is addressed. While some investigators may feel that this is a closed question (the primitives to be extracted are edges) this work illustrates a wide variety of image domain primitives which may be employed. Within the field of machine learning the tension between general learning procedures and domain specific learning procedures is examined. The work characterizes coarse and fine grained knowledge, and argues that machine learning is needed to generate fine grained knowledge. This implies that machine learning will be necessary in such problem areas as computer vision and speech recognition where the input to the system arrives in fine grained (signal domain) format. At the same time the learning system must employ domain specific knowledge supplied to it by the system designer in coarse grained format. The work shows how such coarse grained domain specific knowledge can be incorporated into systems employing a general learning strategy.